With foundation models running out of data for training, companies are placing their bets on the use of synthetic data for training the models. But based on a recent research paper, “indiscriminate use of model-generated content in training causes...
With foundation models running out of data for training, companies are placing their bets on the use of synthetic data for training the models. When Meta released the latest iteration of its open-source model, Llama 3.1 405B, it also updated the model’s licence to generate synthetic data which can be further used to train proprietary small models. On paper, this sounds plausible.
But recent studies cast doubts on this model. Based on a research paper published in science journal ‘Nature’, “indiscriminate use of model-generated content in training causes irreversible defects in the resulting models”.
Meta partners with Reliance to lease 168 MW AI data centre in Jamnagar
